Abstract: To address the nonlinear and highly complex modeling challenges in predicting water inrush disasters from Ordovician limestone confined aquifers during deep coal mining, this study develops an intelligent prediction model (BO-XGBoost) integrating Bayesian Optimization (BO) and eXtreme Gradient Boosting (XGBoost). The research aims to enhance prediction accuracy and model interpretability, covering the modeling of nonlinear hydrogeological responses, identification of key hazard-inducing factors, and spatial risk visualization. The BO algorithm is employed to globally optimize the hyperparameters of XGBoost, overcoming inefficiency and local optima issues inherent in traditional parameter tuning methods. The model iteratively generates Classification and Regression Trees (CART), outputs log-odds ratios, and through regularization and weighted mapping, transforms the results into water inrush probabilities within the 0,1 range, enabling high-precision classification. These probability outputs are then integrated into a Geographic Information System (GIS) platform for spatial visualization, generating refined risk distribution maps. Experimental results show that the model achieves a classification accuracy of 85% and an AUC value of 0.83, significantly outperforming comparative algorithms such as BO-GBDT, BO-RF, and standard XGBoost. Application and validation at Sima Coal Mine demonstrate a spatial structural similarity (SSIM) of 71% between the predicted water inrush probability field and the evaluation results from the traditional water inrush coefficient method. Risk levels are adaptively classified using probabilistic thresholds, yielding results with greater spatial detail and differentiation. SHAP (Shapley Additive Explanations) based interpretability analysis reveals that the geological structure defect index and the equivalent thickness of the aquiclude are the dominant factors influencing water inrush, with feature importances of 32.7% and 28.4%, respectively, consistent with rock mass fracture seepage theory. The developed BO-XGBoost-GIS collaborative model enables high-precision quantitative prediction and mechanistic interpretability of water inrush risk, providing an effective technical approach for water inrush prevention and control in deep coal mines with Ordovician aquifers.